Salvaged premise from #67065 (@webtecnica, issue #67027), reimplemented:
get_env_value() read os.environ first with no secret-scope check, so a
multiplexed profile turn could serve another profile's credential. Its
siblings get_env_value_prefer_dotenv and gateway.config._getenv were
already scope-aware.
Reimplementation note: the original diff called get_secret() but fell
through to os.environ on a scoped miss — re-opening the exact leak it
targeted (flagged by the sweeper review). This version delegates policy
fully to agent.secret_scope.get_secret (global vars pass through; scope
authoritative under multiplexing; legacy environ behavior when off;
UnscopedSecretError propagates fail-closed), then falls back to .env.
6 regression tests incl. the #67027 repro (envless profile + multiplexed
turn -> None, not the other profile's key); sabotage-verified RED on the
old implementation.
Salvaged from #51604 (@JoaoMarcos44, issue #51603): resolve_anthropic_token()
and run_oauth_setup_token() in agent/anthropic_adapter.py read
ANTHROPIC_TOKEN / CLAUDE_CODE_OAUTH_TOKEN / ANTHROPIC_API_KEY via bare
os.getenv(), bypassing agent.secret_scope — a cross-profile over-read in
multiplex mode. Every other provider routes through
runtime_provider._getenv -> get_secret; the adapter now does the same via
a local _getenv wrapper (identical to os.getenv when multiplexing is off,
scope-authoritative + fail-closed when on).
Dropped from the original PR: the cron scheduler hunks (superseded by
fdab380a1a which installs the per-job profile scope) and the unrelated
hermes_logging Windows hunk (scope creep).
Includes the PR's RED->GREEN scope-isolation test file (6 tests).
The salvaged cleanup (#75197) scrubbed every known Hermes key absent from
the profile .env — deleting user-shell-exported credentials
(export OPENAI_API_KEY=...) on every hermes invocation, a documented flow
the author's own failing test_dump_flags_shell_only_key_not_in_dotenv
confirmed. A child process cannot distinguish shell exports from
parent-process leakage, so the scrub now covers ONLY
_PROFILE_MANAGED_ENV_KEYS (ACP routing keys: HERMES_ACP_*,
HERMES_COPILOT_ACP_*, COPILOT_CLI_PATH, COPILOT_ACP_BASE_URL) —
the vector from #75141. Cross-profile credential isolation is owned at
read time by agent.secret_scope.get_secret.
Adds shell-export survival regression + a scope-invariant test that fails
if the scrub set is ever widened toward credential-shaped keys.
Align load_hermes_dotenv() with reload_env() so known Hermes env vars
absent from the active profile .env are removed from os.environ instead
of leaking from a parent process / other profile.
Register ACP-related keys (HERMES_ACP_AUTH_METHOD, HERMES_COPILOT_ACP_*,
COPILOT_CLI_PATH, COPILOT_ACP_BASE_URL) in _EXTRA_ENV_KEYS so they
participate in known-key cleanup.
This is the same isolation gap class as #68367 / #66930, but:
- Not Desktop-only spawn scrub — CLI/gateway restart inheritance
- Not Matrix/messaging auto-enable only — copilot-acp provider/ACP config
- Startup dotenv clear so *any* inheritance path is covered
Example: HERMES_ACP_AUTH_METHOD=cursor_login leaking into a Claude Code
ACP profile caused authenticate -> Internal error -> Discord
'model provider failed after retries'.
The npm 12 requirement (f88ed6c717) strands every system-Node install:
no shipping Node bundles npm >=12, engine-strict makes EBADENGINE fatal,
and the recovery in npm_engine.py refuses to touch a system npm — so
'hermes update' leaves the install in a mixed state (updated code, stale
Node deps, no TUI/web/desktop rebuild) with only a manual-fix hint.
Instead of modifying the user's toolchain (still never done), the
EBADENGINE recovery now provisions Hermes' own managed Node tree under
$HERMES_HOME/node — the same pinned-nodejs.org path install.sh and
install.ps1 use — upgrades THAT npm into the required range, and hands
the caller the managed npm for its single retry.
- hermes_constants.bootstrap_hermes_managed_node(): cross-platform
provisioning (POSIX via node-bootstrap.sh _nb_install_bundled_node,
Windows via the existing portable-zip download); reuses a healthy tree.
- node-bootstrap.sh: HERMES_NODE_SKIP_LINKS=1 skips the ~/.local/bin
node/npm/npx symlinks so the private tree never shadows the user's
own toolchain on PATH.
- maybe_repair_npm_engine() now returns the npm path to retry with
(managed-in-place upgrade or freshly provisioned runtime); both call
sites retry with the returned path and put the managed tree first on
PATH so npm lifecycle scripts resolve the managed node.
- Node-only mismatches on a foreign npm are now also recoverable (the
managed tree ships a supported Node); on a managed npm they still
correctly decline.
E2E (real download, temp HERMES_HOME): provisioned node v22.23.2,
upgraded bundled npm 10.9.4 -> 12.0.2, system npm byte-identical after,
no ~/.local/bin links re-pointed, healthy-tree reuse in 0.05s.
Three lower-severity core-tool robustness fixes from a targeted audit, each
reproduced live:
1. terminal_tool did not validate non-positive timeouts. 'timeout or default'
silently coerced 0 to the config default (0 can't mean 'no timeout'), and a
negative value is truthy so it flowed into 'deadline = now + timeout' and
fired an immediate '-Ns' timeout. Reject timeout <= 0 with a clear message.
2. fuzzy_find_and_replace accepted a whitespace-only old_string, which matches
trivially (blank line / run of spaces) and mass-replaces under replace_all
or raises an opaque ambiguity error. Reject it alongside the empty check.
3. The '/private/var/' sensitive-path prefix over-blocked ALL macOS temp-file
writes: , /tmp, and /var/folders realpath into /private/var/folders
on macOS (and paths are resolved through symlinks), and /private/var/tmp is
a normal temp dir. Narrowed to the genuinely-sensitive subtrees
(/private/var/db, /private/var/root); /etc and /private/etc stay blocked.
All verified with sabotage-checked regression tests. 85 terminal/fuzzy/file
tests pass; normal timeouts, legit replacements, and /var + /boot + /etc
blocking are unaffected.
Two V4A parse/validate bugs found in a core-tools audit, reproduced live:
1. CRLF patch body injected stray carriage returns. parse_v4a_patch split
on '\n' only, so a CRLF-encoded patch kept '\r' inside every HunkLine
and wrote mixed line endings into an LF file; the anchored Begin/End
markers could also fail to match because of the trailing '\r'. Strip a
trailing '\r' from each line at split time.
2. Move-then-Update of the same file was rejected. _validate_operations read
the UPDATE target from disk before the MOVE ran, so 'Move a->b' + 'Update
b' failed validation with 'b: file not found'. Added a small pending-move
overlay so UPDATE/DELETE/MOVE reads during validation see prior ops'
effects (moved-in destinations resolve, moved-away sources read as gone),
while a genuine 'destination already exists' conflict is still caught.
Both verified with sabotage-checked regression tests. 113 patch/fuzzy/file
tests pass.
Strategy 9 (context_aware, the last-resort fuzzy strategy used by
patch_replace, V4A UPDATE hunks, and skill_manage) had two serious flaws,
both reproduced live against current main:
1. CORRECTNESS: it accepted a block when >=50% of its lines were >=0.80
similar. A 2-line pattern with one real line and one garbage line matched,
silently deleting the non-matching line and persisting a wrong edit as
success. Now requires the first AND last lines to anchor-match and EVERY
non-blank pattern line to be >=0.80 similar — one garbage line disqualifies
the block.
2. PERFORMANCE: it scored every content window with per-line SequenceMatcher,
so every failed match paid O(file_lines x pattern_lines) — measured ~5.5s
for a single 40-line no-match on a 10k-line file, per hunk. The first/last
line anchor pre-filter skips non-candidate windows: same case now ~160ms
(34x faster).
Also gate replace_all: a similarity-based strategy (block_anchor,
context_aware) with multiple matches under replace_all would overwrite every
approximate block, not just exact ones. Now refused with a clear error
directing the caller to precise text.
All verified with sabotage-checked regression tests (fail against the old
50% logic). 158 file/patch/fuzzy tests pass; legit fuzzy edits (indent drift,
unique near-match) unaffected.
Two DATA-LOSS bugs in ShellFileOperations found in a core-tools audit,
each reproduced live against current main:
1. Non-UTF-8 file content silently corrupted on read->write. The terminal
env decodes stdout with errors='replace', so a latin-1/8859 file's bytes
arrive as U+FFFD before _is_likely_binary inspects them. U+FFFD is
'printable', so the >30%-non-printable check never flagged it, and the
agent would read the mojibake and write it back, permanently replacing the
original bytes. Fix: treat a sample containing U+FFFD as binary (read-only).
2. Writing through a symlink destroyed the link and orphaned the target. The
atomic temp-file + 'mv -f' swap replaced the symlink itself with a plain
file; the real target was never updated. Fix: resolve the link with
readlink -f/realpath first and recompute the temp dir from the resolved
target so the mv stays same-filesystem atomic. Broken links fall back to
the original path (no regression).
Both verified with sabotage-checked regression tests (fail without the fix).
Proper UTF-8 text (incl. non-ASCII) and plain-file writes are unaffected.
Two independent HIGH-severity correctness bugs found in a core-tools audit,
each reproduced live against current main:
1. Read-dedup was never evicted after a write on non-default tasks.
_invalidate_dedup_for_path looked up the read-tracker under the correct
task_id but resolved the path with _resolve_path(filepath) — which
DEFAULTS task_id='default'. The dedup cache is keyed by the task-resolved
absolute path, so for any task whose workspace cwd differs from the process
cwd (every -w worktree / Desktop / ACP session using relative paths) the
computed key never matched and the stale entry was never removed. A
read_file after a write_file/patch could then return the OLD content stub
when mtime coincided. Fix: pass task_id through.
2. A per-command workdir override permanently hijacked the session cwd.
The post-command dual-write unconditionally recorded env.cwd (stamped to
the transient workdir) into the durable session-cwd store, so every later
command that omitted workdir inherited the one-off directory — contradicting
the documented 'Working directory for this command' contract. Fix: skip the
session-cwd record when workdir was explicitly supplied.
Both verified with sabotage-checked regression tests (fail without the fix).
Users following abbreviated links guess /docs/quickstart and
/docs/installation and hit raw GitHub-Pages 404s — the real pages live
under /docs/getting-started/. Add client redirects for both.
Consumer-onboarding audit finding #1, Aug 2026.
Cancelling the API-key prompt mid-wizard (Enter → 'Cancelled.') let the
wizard continue through Terminal/Gateway/Tools and finish 'successfully'
with no model configured — the user exits believing they're set up, then
hits a broken chat.
_print_setup_summary() (called by every setup path: full, quick,
blank-slate, portal) now probes resolve_provider() and, when nothing is
configured, prints an unmissable warning with the two one-line fixes
(hermes model / hermes setup --portal).
Consumer-onboarding audit finding #7 (sev 4), Aug 2026.
Teknium review on #62947: os.environ.clear()/update around deferred
loaders is unsafe under concurrency and misses teams_pipeline's direct
adapter import.
Defer microsoft_teams binding in the Teams adapter, no-op
dotenv.load_dotenv while the SDK imports, keep api_server explicit
disable, and add SDK-import + load_gateway_config canaries.
Fixes#62935
microsoft-teams-apps calls load_dotenv(find_dotenv(usecwd=True)) at import time, which can pull a root-profile .env into every gateway process during plugin_entries() discovery and break profile secret isolation.
Snapshot/restore os.environ around deferred loaders, and honor explicit api_server enabled:false the same way _enable_from_env does for other platforms.
Fixes#62935
A completely unconfigured install previously booted into a working-looking
chat (banner showed model 'unknown'), accepted a message, spun ~30s, then
failed with 'Set OPENROUTER_API_KEY' — a provider the user never chose —
and never offered setup.
- HermesCLI.run() now probes provider readiness at startup (TTY only) and
offers the shared provider picker (hermes model flow, which fronts Quick
Setup / Nous Portal OAuth) when nothing is configured. Decline is
respected; picker state re-syncs into the live CLI so the next turn works
without a restart.
- New silent probe _runtime_credentials_ready(): no printing, no state
mutation; handles keyless local endpoints and callable bearer providers.
- The empty-api-key error is provider-aware: names the actual resolved
provider and points at 'hermes model' / 'hermes setup' instead of
hardcoding OPENROUTER_API_KEY.
- Banner: unconfigured installs render 'no model configured — run /model'
in red instead of the silent 'unknown' model slug.
Consumer-onboarding audit finding #2 (sev 5), Aug 2026.
After the test-suite prune, the Python slices (~2.3m each) are no longer
CI's critical path — Desktop E2E (5.2m, the longest job) and the Docker
build are, and both run on every python-lane PR even when the diff never
leaves tests/. Neither consumes the test suite: Playwright drives the
built app + hermes serve backend, and the image copies installed code.
New python_prod lane = python minus tests-only diffs. e2e-desktop and
docker gate on it; every pytest/lint lane keeps gating on python.
Fail-open contract preserved: .github/ changes and empty diffs set
python_prod=true, and runner infrastructure (scripts/run_tests.sh,
run_tests_parallel.py) is deliberately NOT tests-only since a bad
runner edit can mask real failures.
Replay over the last 231 main commits: 39 (17%) would skip both jobs,
cutting their critical path from ~8m to ~3m. E2E-verified through the
real script entrypoint (tests-only/prod/mixed/fail-open) + 83 tests/ci
green.
Input events are deliberately skipped for the duration of an IME composition
(they carry uncommitted preedit text), so nothing clears the empty marker
until compositionend — the hint kept painting behind the hiragana the user was
composing. Drop the marker as composition starts; the normalizer restores it
if composition ends with nothing committed.
Taking the hint out of the text flow fixed the displacement half of #75960 on
its own — preedit now starts at the field's left edge either way — but the
overlap needed this too.
Co-authored-by: Ryuichi Natori <to-na@users.noreply.github.com>
The ACP prompt path called set_session_vars(session_key=session_id, ...)
without passing session_id, so the HERMES_SESSION_ID ContextVar was bound
to its explicit "" default. Once the session-context machinery is engaged,
_inject_session_context_env treats an explicitly-bound "" as authoritative
and writes it to the child env — so subprocesses spawned during an ACP turn
got an empty HERMES_SESSION_ID instead of the session's own id.
Pass session_id through so child subprocesses carry the correct id.
Salvage of the ACP half of #53454 by @necoweb3 (the V4A-path half is
salvaged separately in the file-tools PR).
Co-authored-by: necoweb3 <sswdarius@gmail.com>
_normalize_managed_eol isolated line-ending churn from real edits by
diffing twice: all dirty files minus files still dirty under
--ignore-cr-at-eol. But 'git diff --name-only --ignore-cr-at-eol'
computes its file list from blob/stat differences BEFORE the CR filter
is applied, so it still lists CR-only files. On git 2.48.1 the two
name-only sets are therefore identical, _eol_only() is always empty, and
a managed Windows checkout gets pinned to core.autocrlf=false with the
whole CRLF tree left dirty — breaking the next 'git checkout' on update
(the exact failure this function exists to prevent).
Compute the real-edit set with 'git diff --numstat --ignore-cr-at-eol'
instead: numstat honors the CR filter (a CR-only file produces no
record), so eol-only files are correctly identified and cleared while
genuine edits are preserved. Pin core.quotepath=false so non-ASCII paths
parse. Verified at 1200 files: 1199 eol-only normalized, one real edit
preserved, autocrlf pinned only after the tree reads clean.
This was a pre-existing failure on main (test_update_eol_churn's
test_churn_across_more_files_than_fit_in_one_argv failed deterministically
on git 2.48.1), surfaced while landing unrelated file-tools PRs.
The remaining /model picker stall after the Copilot backoff fix: whenever
the 1h provider-models disk cache TTL (or the remote model-catalog manifest
TTL) lapsed mid-session, the next picker open blocked on 8-9 serial
/v1/models round-trips (~2-3s measured) plus the catalog manifest fetch
before rendering anything.
Model catalogs change on release timescales, not hourly — so both caches
now use stale-while-revalidate:
- cached_provider_model_ids(): an expired entry whose credential
fingerprint still matches is served immediately; a deduped daemon thread
re-fetches the live catalog and rewrites the disk cache for the next
open. Entries older than 7 days still block on a live fetch, credential
rotation still busts the entry, and force_refresh still bypasses SWR.
- model_catalog.get_catalog(): an expired disk manifest is served
immediately with an off-thread refresh; only a truly cold cache (no disk
copy) blocks on the network.
Measured picker payload build with deliberately-expired caches:
2.9s -> 0.93s (first open in process) / 0.06s (subsequent opens).
Combined with the Copilot fix (#76386): 7.3s -> ~0.06s for the common case.
The empty-composer prompt was painted with an inline `::before`, which puts a
real box in the contenteditable's text flow. Click an empty composer and the
caret lands past the hint instead of at the field's left edge, and a hint that
wraps — a narrow composer, a long locale string — makes the empty composer two
lines tall. It is a hint, so it now sits out of flow: absolutely positioned,
clipped to one line, unselectable and untouchable by the pointer. Measured in
Chromium against the built stylesheet, the caret lands at the left edge for a
wide composer, a narrow one, and a Japanese hint alike.
Backspace on a fresh `@folder:` chip had a second, related problem. Committing
a completion empties the typed token's text node rather than removing it, and
`Range.insertNode` splits the line around the caret, so the chip ends up
between zero-length text nodes. Those read as content: the atomic chip-delete
declined, Chromium's own backspace bounced between the leftovers, and the chip
took extra presses to remove — leaving a `"\n"` draft with the hint still
hidden behind it. Emptiness, the chip-delete, and the DOM normalizer now all
step over that litter.
The stylesheet owns the rule now that it needs `position: relative` on the
editor, so the utility-class constant both composers imported is gone.
patch_tool resolved V4A header paths against the task workspace for
locking, staleness, and reporting, but handed the original (often
relative) patch text to file_ops.patch_v4a — which re-resolved headers
against the backend env's own cwd. When the two diverge (the git-worktree
cwd bug), a relative header landed in a different directory than
everything the tool locked and reported: a silent wrong-file write.
Rewrite Update/Add/Delete/Move File headers to the resolved absolute
paths before apply, only for host-filesystem backends (container/remote
namespaces keep their own paths). Header patterns mirror patch_parser
(no-space ***Update File: form) and cover Move File: src -> dst.
Salvage of #53176 by @necoweb3, reimplemented onto current main (the
original branch predates the sensitive-path/Move-header extraction and
per-path locking now in patch_tool).
Co-authored-by: necoweb3 <sswdarius@gmail.com>
The boundary scan in parse_v4a_patch used substring matching, so a
content line mentioning "*** End Patch" (docs about the patch format,
nested patch text) truncated the patch, and "*** Begin Patch" in
content reset the start boundary — silently dropping already-parsed
operations while reporting success. Match only whole-line markers at
column 0, preserving the no-space "***Begin Patch" tolerance.
The Desktop client writes the SSH session token under $HOME/.hermes/desktop-ssh
(a literal ~/.hermes/desktop-ssh in apps/desktop/electron/remote-lifecycle.ts,
expanded against the account's $HOME), independent of HERMES_HOME and the active
profile. But _read_ssh_session_token_file validated it against
get_hermes_home()/desktop-ssh, which a non-default sticky profile re-homes to
<root>/profiles/<name>/desktop-ssh (and any custom HERMES_HOME points elsewhere).
relative_to() then rejects every token as "not under the desktop-ssh directory",
so SSH remote mode is broken under any non-default profile.
Anchor to Path.home()/.hermes/desktop-ssh so the validator matches the exact
directory the client writes to, across default, profile, and Docker layouts.
Adds profile / custom-root acceptance tests and a profile-local rejection test.
Fixes#69551.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The desktop lint rule added after the original PR bans mirroring reactive
values into refs via useEffect. Rework the setting to seed from draft-null
state, guard profile switches by stale-config identity, and derive the
save base + rollback value from the shared config record instead of
latestConfigRef/lastSavedRef/seededRef.
Adds a new config option terminal.font_family that lets users customize the
CSS font-family for the desktop app's embedded xterm.js terminal.
Previously the font was hardcoded in use-terminal-session.ts:
'JetBrains Mono', 'Cascadia Code', 'SF Mono', Menlo, Consolas, monospace
Now the value from config.yaml (terminal.font_family) is threaded through:
useHermesConfig → PersistentTerminal → TerminalTab → useTerminalSession
When font_family is empty or unset (default), the built-in fallback is used,
preserving backward compatibility. Users with Nerd Fonts installed (e.g.
CaskaydiaCoveNerdFont) can now set:
terminal:
font_family: 'CaskaydiaCoveNerdFont', 'JetBrains Mono', monospace
Closes: #terminal-font-config
The no-args /model picker calls list_authenticated_providers(), which walks
every provider through load_pool(). For copilot, _seed_from_singletons()
re-runs the raw-token -> API-token exchange on every pass. When the exchange
is rejected (HTTP 403: token not Copilot-entitled, revoked, org-blocked),
the transient-network retry loop slept ~4.5s (1.5s + 3.0s backoff) before
degrading to the raw token — and nothing cached the failure, so EVERY picker
open, provider discovery pass, delegation spawn, and dashboard credential
listing paid the full 4.5s again.
Measured on a machine with a 403-rejected gh token: /model picker payload
build went from 7.3s to 1.0s cold and 0.06s warm.
Fixes:
- Permanent HTTP rejections (401/403/404) skip the retry backoff entirely —
the loop exists for startup network races, not auth rejections.
- Negative cache keyed on token fingerprint: failed exchanges are not
re-attempted for 30min (auth rejection) / 60s (transient network error).
- Success and evict_cached_exchanged_token() both clear the negative-cache
entry, so the runtime stale-credential recovery path still forces a fresh
exchange.
Two regressions from moving Edit Models into the shared catalog:
The row opened its own separator group, so the composer showed two rules in
the trailing block where it had always shown one. Edit Models now renders
inside the host footer's group — same single separator, same order.
The panel also read the catalog with a non-reactive getQueryData. With no
model in the session store yet (a fresh draft), currentPickerSelection falls
back to the catalog's reported current, and a cache peek never repaints once
that resolves — the menu could sit on a stale or empty selection. Back to a
live useQuery on the same key, which React Query dedupes against the menu's
own subscription rather than double-fetching.
The board listed every model while the composer honoured the user's Edit
Models shortlist, because visibility arrived as a prop each caller opted into.
Curation is one stored preference, so a per-caller opt-in guarantees the two
surfaces eventually disagree about what "my models" means — the exact drift
extracting the menu was meant to end.
ModelCatalogMenu now reads $visibleModels itself and renders the Edit Models
row, so every picker shows the same shortlist and offers the same way to
change it. The composer's footer keeps only Refresh Models. No plugin surface
is involved: the menu is core and the plugin just mounts it.
Also restores the session-scoped catalog fetch the extraction dropped. The
composer's query key must carry its session id — the app invalidates the
session-scoped key on model changes, so a global key would have gone stale
mid-conversation.
The desktop board had no model control at all: a task ran whatever the profile
you assigned it to happened to be configured with, and the only way to point
one task elsewhere was `hermes kanban set-model` or the browser dashboard's
flat provider:model select.
Adds a Model row to New Task and to the task drawer, rendering the composer's
own picker via the SDK — same search, same provider groups, same submenu — so
the board and the chat bar cannot drift. Unset reads "Profile default" and
changes nothing; a pin reads "provider: model · High" with an inline clear.
Presets are read-only here: picking a model seeds the depth from what you last
used for it, but a per-task choice never rewrites what the composer opens at.
Fast mode is omitted rather than shown-and-ignored — it's a live-session
request parameter with no worker-spawn equivalent.
The New Task dialog opts out of DialogContent's clip: the dialog publishes
itself as the portal container for popovers opened inside it, so its
overflow-y-auto cropped the menu at the dialog's edge. The general fix is in
flight as #75600; this override is scoped to one dialog to avoid conflicting
with it and disappears when that lands.
The model menu — search, provider grouping, -fast family collapse, keyboard
selection, the per-row thinking/effort submenu — was welded to the chat
composer's session writes, so any other surface wanting a model picker had to
fork it and drift.
Splits rendering from meaning. ModelCatalogMenu owns the catalog and the
navigation; a ModelMenuController decides what a selection DOES. The composer
is now one controller over it, keeping its session scoping, sticky manual
pick, preset restore, MoA presets, and rollback-on-failed-write intact.
ModelEditSubmenu becomes pure: it reports edits instead of performing them.
It previously called setCurrentReasoningEffort and config.set inline, so any
non-composer host would have silently retargeted the user's live chat when
they picked an effort. Its default effort is passed in rather than read from a
store, which is what lets it render outside a session at all.
Exported through the SDK so plugins consume the real component instead of a
copy. The composer's existing behaviour suite passes unchanged against it.
Carries the new column through create, PATCH, and bulk. Clearing is an
explicit clear_reasoning_effort flag rather than a null, because a null in a
PATCH body means "field not sent", not "set to NULL" — the same shape the
model override already uses, and the reason "none" can stay a real value.
Tests cover normalization, the depth-survives-a-model-clear invariant, both
spawn-argv branches, and the REST round-trip. One asserts the worker CLI
actually accepts the --reasoning flag the dispatcher emits: a spawn arg no
parser accepts would fail every dispatch while every unit test stayed green.
A task could already pin a model and provider, but not how hard the worker
thinks: reasoning effort came from the assigned profile's config and nothing
per-task could reach it. Pairing a small model with high effort, or a big one
with thinking off, meant editing the worker profile itself.
Adds a tasks.reasoning_effort column (migrated, NULL = inherit the profile)
with set_reasoning_effort(), a create_task kwarg, and a --reasoning spawn flag.
Kept deliberately independent of model_override: a task may run the profile's
own model at a different depth, and clearing a model override no longer resets
the depth the operator chose. "none" is a value (thinking off), not a clear.
--reasoning is new on the CLI too — the level was only reachable through the
/reasoning slash command, so the dispatcher had no flag to pass. It overrides
agent.reasoning_effort for one run and is never persisted.
The micro-action strip owned this skin inline: a full-radius hairline pill
on the composer's own fill behind a blur, sized to `--composer-control-size`.
It's the right look for anything that floats over the composer, so lift it
into `composer-dock` next to the other shared composer surfaces and have the
strip compose it with its own width cap and disabled state.
Teknium